A forecast is only useful if someone changes a decision because of it.
In brief
- Forecasting works when it is tied to a decision: stock, staffing, targets or cash.
- Data quality and definitions matter more than the model. Start with a simple baseline and beat it.
- Test against history the model has not seen, and report the range of error, not only the number.
- Build the forecast into the weekly or monthly planning routine, with an owner.
This article describes a method. It does not describe a delivered model.
Many companies want an AI sales forecast. Few can say what they would do differently with one. The projects that work begin with the decision, not the algorithm.
The real issue is not accuracy. It is use.
A slightly less accurate forecast that planners trust and use beats a clever one that nobody believes. Trust comes from clear data, a transparent baseline and honest error ranges.
From data to action
| Stage | Question | What good looks like |
|---|---|---|
| Decision | What will we do with the forecast? | A named decision, owner and time frame |
| Data | Do we have history we trust? | Sales, prices, promotions and stock in one place with agreed definitions |
| Baseline | What does a simple method give? | A transparent benchmark such as last year’s pattern |
| Model | Does an AI method improve on the baseline? | Better results on data the model has not seen |
| Error | How wrong can it be? | A stated range, by product or region |
| Routine | How is it used? | Reviewed in planning meetings and compared with actuals |
Building it step by step
1. Start with the decision
“Order the right stock for next month” is a decision. “Forecast sales” is not. The decision sets the time frame and the level of detail.
2. Prepare the data
Combine sales history with what explains it: prices, promotions, seasonality, events, stock-outs. Agree what counts as a sale and fix gaps. This often takes most of the effort.
3. Set a simple baseline
Every model must be compared with something simple. If it cannot beat last year’s pattern, it is not ready.
4. Try the model and test it fairly
Train on earlier periods and test on later ones. Never judge a model on data it has already seen.
5. Report the error
Give a range, and show where the model is weak, such as new products or promotions. Planners can handle uncertainty if it is stated.
6. Put it in the routine
Review each forecast against actual results, learn from the gaps and keep a person responsible.
Five mistakes we see repeatedly
- No decision attached.
- Messy or inconsistent history.
- Skipping the baseline.
- Testing on training data.
- Presenting a single number with no range.
Our experience
Forecasting depends on trusted, connected data, and building that foundation is where much of our work lies. We are not describing a delivered forecasting model here. This is a method drawn from how data projects tend to go.
We can explore your data and the decision a forecast would support.
What decision would you change if you had a better forecast?
A forecast people use
Start from the decision, prepare the data, beat a simple baseline, state the error, and make the forecast part of planning. The model is the smallest part of the work.
Talk to Balerion about preparing your data and planning a forecasting pilot. Email us at talk@balerion.co.th.
